Write-up by Helena Ward, Graduate Scholar and DPhil candidate in Philosophy.
On Thursday 12 June, the Institute for Ethics in AI welcomed Professor Jessica Riskin, Historian of Science, for the Fourth Annual Lecture of the Institute for Ethics in AI. Hosted by Professor Sir Nigel Shadbolt, Riskin’s lecture examined the implications of what she calls ‘the onion axiom’ for the history of artificial intelligence.
What is the Onion Axiom?
Ever since the founding moments of artificial intelligence in the mid-twentieth century, the principal researchers and writers on the subject—from Alan Turing and Norbert Weiner to Rodney Brooks and Daniel Dennet—have adopted a behaviourist and reductive assumption: the onion axiom.
The onion axiom states that intelligence is nothing more than its appearance. That if you could peel away all the layers of an onion, at the core you would find nothing but emptiness: there’s no one in there.
Think of our interactions with other people. When you and I converse with one another, we reciprocally project our imaginations to one another’s personhood—I imagine what your motivations, suppositions, and intentions are, and you mine. Now think of our interactions with an LLM, or even the mere observation of a mechanical machine—we might similarly project intelligence and intentions onto an artefact. This time, however, there’s no one in there.
The Intellectual World From Which the Emergence of Artificial Intelligence Took Place
The idea that machines can imitate intelligence has appeared at numerous points and forms in history. Current machines are capable of performing tasks we once thought were unique to human intelligence—from playing chess and writing poetry, to learning and providing emotional support.
The various tasks that artificial machines are capable of, would, if performed by humans or non-human animals, be taken as signs of an intelligent self.
From the fact that machines could appear intelligent from the outside but have no intelligence within, many early pioneers of artificial intelligence came to adopt a reductionist view of intelligence. From around the middle of the 20th century, it became magical and mystical to believe in intelligence; researchers started to believe that intelligence—whether in humans or machines—did not exist. Intelligence was an illusion, a mirage, a trick of light and colour.
Riskin illustrates the prevalence of the onion axiom through historical analysis, we consider a few such examples below:
Elmer and Elsie
One of the earliest examples was Elmer and Elsie, two electromechanical robots built in the late 1940s by William Walter. Elmer and Elsie were described as the first robots in history that were programmed to ‘think’. They were restless explorers, navigating the environment and by way of photocell receptors—they were attracted to moderate light sources, and repelled by bright ones. When interacting with their own image in a mirror, or looking at one another, the tortoises appeared to be self-aware, and aware of another. They would first approach one another, or oneself, then back away, and inquire again—jigging back and forth with seeming curiosity.
These appearances were explained away by the fact that the tortoises had a headlamp placed on their heads, which would turn off automatically when the photocell was exposed to bright light. The robot was drawn to its own reflection, or the other, (given that they were attracted to moderate light sources), but, too close, and they backed away (given that they were repelled by bright light sources).
As Walter argued, were this behaviour to be exhibited in animals, it might be accepted as evidence of intelligence, self-awareness or mutual recognition. But once we understand the mechanisms by which these tortoises behave, the intelligence evaporates—their self-hood thus becomes plausible only if we equate appearance with reality.
Herbert
Another example is an embodied robot called Herbert, designed by Rodney Brookes. Herbert was a machine that would wonder around the office stealing empty soda cans. From the outside, Herbet appears to be intelligent and have intentions, but his intelligence is entirely mechanistic. Intelligence is, Brookes insists, merely in the eye of the beholder.
Searle’s Chinese Room Experiment
A final example is Searle’s Chinese room experiment. To those unfamiliar, Searle asks us to imagine a person in a room who does not understand Chinese, receiving Chinese characters through a slot in the wall. Using a rule book written in English, they match symbols to produce correct Chinese responses. To an outsider, it appears as though the person on the inside understands Chinese, even though they do not. Once again, we see appearance failing to match up to reality.
So, as Riskin examines, during the early emergence of artificial intelligence, many believed that intelligence was an illusion, that, once we know the rules a machine follows, how it works, and what is going on in the inside, the onion axiom rears its empty centre. A science of the inner workings’ intelligence became impossible, even paradoxical, since once we manage to explain the appearance of intelligence, it evaporates. As Ross Ashby puts it:
‘real intelligence does not exist. It is a myth. It has come into existence in the same way that the ideal of “real” magic comes to a child who sees conjuring tricks”
What is an Intelligent Machine (1961: 271)
While the machines around us appear to be getting more and more intelligent, we haven’t gotten any better at understanding living intelligence. Why is this? One possibility Riskin raised was that—quite astoundingly—artificial intelligence emerged under the presupposition that intelligence itself did not exist.
Implications—Personalised AIs
One of the practical values of history of science work is that it can provide us with a perspective on current science. What implications might the onion axiom have for thinking about current forms of artificial intelligence?
One central tenet of the Onion Axiom is that artificially intelligent machines can appear to have an individual self, even though there is no ‘I’ in there. This is particularly relevant for thinking about personalised LLMs.
From chat-bots which imitate lost-loved ones, to AI friends, and therapists, personalised models respond to user prompts while adopting what seems like an individual character or self. By way of emotive language, expressions of an ‘I’, set personality traits and backgrounds, these models exhibit astoundingly sensical responses, often indistinguishable from human interlocutors. Such models appear, from the outside, to arise from an individual self, but—as the onion axiom states—there’s no ‘I’ in there.
One lesson from the onion axiom might be that we should adopt guardrails to minimise the deceptiveness of these models. We should make explicit to users that there really is no one in there—to peel back the layers of the onion—and clearly demarcate the appearance of a self, from the reality of one.
Annual Lecture:
- Discover the biographies of our annual lecture speaker and host
- See the Recording of the annual lecture